When Translation Is Not Transition: Target-Path Placement After Model-Visible Information in Governed LLM Revision

LLM revision systems can receive the information needed for a required update and still fail to place that information in the required structured location or preserve unrelated state. This paper studies that distinction in a matched 48-case benchmark with deterministic rules specifying what must change, what may change, and what must remain unchanged. We call successful placement under those rules target-path fidelity. In the historical evaluation, Arms A and B each accumulated 56 deterministic rule violations, and machine-rated quality differed by -0.03125 for B relative to A (paired-bootstrap 95% CI [-0.2604167, +0.1875]). Among 11 cases missing a required transition at the initial stage, 4 were exactly repaired after revision, 7 remained unresolved, and 3 also lost a constraint prohibiting external action. A minimal placement instruction then passed 103/103 selected software checks, but a follow-up 48-case evaluation remained mixed: rule-violation totals were A=55 and B=53, paired zero-error discordants were 4 versus 3 (exact McNemar p=1.0), and machine-rated quality B-A was +0.086806 (paired-bootstrap 95% CI [-0.184028, +0.347222]). A later evaluation with stronger actionable guidance sharpened the negative result. Re-evaluating all 96 arm-level revised records from the designated metric field produced rule-violation totals A=52 and B=59; the inherited summary A=55/B=53 agreed with only 70/96 current records. Across all 48 cases, structural-error totals were A=4 and B=12. A separate nine-case subset selected because the arms ended with different structural outcomes contributed A=1 and B=9 of those totals; within that subset, B showed two required-state regressions and no required-state repairs. Underaction occurred on eight authorized arm rows and on zero restraint arm rows. A treatment-blind machine review of 24 cases, checked and cosigned by the human investigator only after generation, qualitatively converged on the same failure pattern. The bounded result is that visible target information, actionable feedback, exact state/path placement, preservation of unrelated state, deterministic validator success, and aggregate outcome are separable estimands.